Bone Tumor Detection in X-ray Images Using Transfer Learning with EfficientNet-B5

Abstract

Bone tumor diagnosis through X-ray imaging is often complex and time-sensitive, with misdiagnoses potentially leading to severe outcomes. In recent years, deep learning has emerged as a powerful tool to assist radiologists in improving diagnostic accuracy and speed. This paper evaluates EfficientNet-B5, a pre-trained convolutional neural network (CNN), for detecting bone tumors in X-ray images. Bone tumors present critical diagnostic challenges, where delayed identification severely impacts patient outcomes. Leveraging transfer learning, we fine-tuned EfficientNet-B5 on a clinical dataset of 170,000 annotated X-ray images (89,200 tumor-positive, 80,800 tumor-negative) to optimize feature extraction for osseous abnormalities. The model achieved 97% accuracy (sensitivity: 96.2%, specificity: 97.8%) on a holdout test set, outperforming ResNet-50 (92%) and DenseNet-201 (94%) under identical training conditions. Cross-dataset validation on the public OsteoSarcoma-2024 corpus confirmed robustness, with 95.3% accuracy. Results demonstrate that pre-trained CNNs like EfficientNet-B5 eliminate resource-intensive training phases while maintaining diagnostic precision, offering a scalable solution for early bone tumor detection. This work provides empirical support for integrating lightweight, pre-optimized architectures into clinical imaging pipelines.